journal.c

linux 内核源代码

	struct reiserfs_journal_list *pjl;
	struct reiserfs_journal_cnode *cn, *last;
	int count;
	int was_jwait = 0;
	int was_dirty = 0;
	struct buffer_head *saved_bh;
	unsigned long j_len_saved = jl->j_len;
	struct reiserfs_journal *journal = SB_JOURNAL(s);
	int err = 0;

	BUG_ON(j_len_saved <= 0);

	if (atomic_read(&journal->j_wcount) != 0) {
		reiserfs_warning(s,
				 "clm-2048: flush_journal_list called with wcount %d",
				 atomic_read(&journal->j_wcount));
	}
	BUG_ON(jl->j_trans_id == 0);

	/* if flushall == 0, the lock is already held */
	if (flushall) {
		down(&journal->j_flush_sem);
	} else if (!down_trylock(&journal->j_flush_sem)) {
		BUG();
	}

	count = 0;
	if (j_len_saved > journal->j_trans_max) {
		reiserfs_panic(s,
			       "journal-715: flush_journal_list, length is %lu, trans id %lu\n",
			       j_len_saved, jl->j_trans_id);
		return 0;
	}

	get_fs_excl();

	/* if all the work is already done, get out of here */
	if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
	    atomic_read(&(jl->j_commit_left)) <= 0) {
		goto flush_older_and_return;
	}

	/* start by putting the commit list on disk.  This will also flush 
	 ** the commit lists of any olders transactions
	 */
	flush_commit_list(s, jl, 1);

	if (!(jl->j_state & LIST_DIRTY)
	    && !reiserfs_is_journal_aborted(journal))
		BUG();

	/* are we done now? */
	if (atomic_read(&(jl->j_nonzerolen)) <= 0 &&
	    atomic_read(&(jl->j_commit_left)) <= 0) {
		goto flush_older_and_return;
	}

	/* loop through each cnode, see if we need to write it, 
	 ** or wait on a more recent transaction, or just ignore it 
	 */
	if (atomic_read(&(journal->j_wcount)) != 0) {
		reiserfs_panic(s,
			       "journal-844: panic journal list is flushing, wcount is not 0\n");
	}
	cn = jl->j_realblock;
	while (cn) {
		was_jwait = 0;
		was_dirty = 0;
		saved_bh = NULL;
		/* blocknr of 0 is no longer in the hash, ignore it */
		if (cn->blocknr == 0) {
			goto free_cnode;
		}

		/* This transaction failed commit. Don't write out to the disk */
		if (!(jl->j_state & LIST_DIRTY))
			goto free_cnode;

		pjl = find_newer_jl_for_cn(cn);
		/* the order is important here.  We check pjl to make sure we
		 ** don't clear BH_JDirty_wait if we aren't the one writing this
		 ** block to disk
		 */
		if (!pjl && cn->bh) {
			saved_bh = cn->bh;

			/* we do this to make sure nobody releases the buffer while 
			 ** we are working with it 
			 */
			get_bh(saved_bh);

			if (buffer_journal_dirty(saved_bh)) {
				BUG_ON(!can_dirty(cn));
				was_jwait = 1;
				was_dirty = 1;
			} else if (can_dirty(cn)) {
				/* everything with !pjl && jwait should be writable */
				BUG();
			}
		}

		/* if someone has this block in a newer transaction, just make
		 ** sure they are committed, and don't try writing it to disk
		 */
		if (pjl) {
			if (atomic_read(&pjl->j_commit_left))
				flush_commit_list(s, pjl, 1);
			goto free_cnode;
		}

		/* bh == NULL when the block got to disk on its own, OR, 
		 ** the block got freed in a future transaction 
		 */
		if (saved_bh == NULL) {
			goto free_cnode;
		}

		/* this should never happen.  kupdate_one_transaction has this list
		 ** locked while it works, so we should never see a buffer here that
		 ** is not marked JDirty_wait
		 */
		if ((!was_jwait) && !buffer_locked(saved_bh)) {
			reiserfs_warning(s,
					 "journal-813: BAD! buffer %llu %cdirty %cjwait, "
					 "not in a newer tranasction",
					 (unsigned long long)saved_bh->
					 b_blocknr, was_dirty ? ' ' : '!',
					 was_jwait ? ' ' : '!');
		}
		if (was_dirty) {
			/* we inc again because saved_bh gets decremented at free_cnode */
			get_bh(saved_bh);
			set_bit(BLOCK_NEEDS_FLUSH, &cn->state);
			lock_buffer(saved_bh);
			BUG_ON(cn->blocknr != saved_bh->b_blocknr);
			if (buffer_dirty(saved_bh))
				submit_logged_buffer(saved_bh);
			else
				unlock_buffer(saved_bh);
			count++;
		} else {
			reiserfs_warning(s,
					 "clm-2082: Unable to flush buffer %llu in %s",
					 (unsigned long long)saved_bh->
					 b_blocknr, __FUNCTION__);
		}
	      free_cnode:
		last = cn;
		cn = cn->next;
		if (saved_bh) {
			/* we incremented this to keep others from taking the buffer head away */
			put_bh(saved_bh);
			if (atomic_read(&(saved_bh->b_count)) < 0) {
				reiserfs_warning(s,
						 "journal-945: saved_bh->b_count < 0");
			}
		}
	}
	if (count > 0) {
		cn = jl->j_realblock;
		while (cn) {
			if (test_bit(BLOCK_NEEDS_FLUSH, &cn->state)) {
				if (!cn->bh) {
					reiserfs_panic(s,
						       "journal-1011: cn->bh is NULL\n");
				}
				wait_on_buffer(cn->bh);
				if (!cn->bh) {
					reiserfs_panic(s,
						       "journal-1012: cn->bh is NULL\n");
				}
				if (unlikely(!buffer_uptodate(cn->bh))) {
#ifdef CONFIG_REISERFS_CHECK
					reiserfs_warning(s,
							 "journal-949: buffer write failed\n");
#endif
					err = -EIO;
				}
				/* note, we must clear the JDirty_wait bit after the up to date
				 ** check, otherwise we race against our flushpage routine
				 */
				BUG_ON(!test_clear_buffer_journal_dirty
				       (cn->bh));

				/* drop one ref for us */
				put_bh(cn->bh);
				/* drop one ref for journal_mark_dirty */
				release_buffer_page(cn->bh);
			}
			cn = cn->next;
		}
	}

	if (err)
		reiserfs_abort(s, -EIO,
			       "Write error while pushing transaction to disk in %s",
			       __FUNCTION__);
      flush_older_and_return:

	/* before we can update the journal header block, we _must_ flush all 
	 ** real blocks from all older transactions to disk.  This is because
	 ** once the header block is updated, this transaction will not be
	 ** replayed after a crash
	 */
	if (flushall) {
		flush_older_journal_lists(s, jl);
	}

	err = journal->j_errno;
	/* before we can remove everything from the hash tables for this 
	 ** transaction, we must make sure it can never be replayed
	 **
	 ** since we are only called from do_journal_end, we know for sure there
	 ** are no allocations going on while we are flushing journal lists.  So,
	 ** we only need to update the journal header block for the last list
	 ** being flushed
	 */
	if (!err && flushall) {
		err =
		    update_journal_header_block(s,
						(jl->j_start + jl->j_len +
						 2) % SB_ONDISK_JOURNAL_SIZE(s),
						jl->j_trans_id);
		if (err)
			reiserfs_abort(s, -EIO,
				       "Write error while updating journal header in %s",
				       __FUNCTION__);
	}
	remove_all_from_journal_list(s, jl, 0);
	list_del_init(&jl->j_list);
	journal->j_num_lists--;
	del_from_work_list(s, jl);

	if (journal->j_last_flush_id != 0 &&
	    (jl->j_trans_id - journal->j_last_flush_id) != 1) {
		reiserfs_warning(s, "clm-2201: last flush %lu, current %lu",
				 journal->j_last_flush_id, jl->j_trans_id);
	}
	journal->j_last_flush_id = jl->j_trans_id;

	/* not strictly required since we are freeing the list, but it should
	 * help find code using dead lists later on
	 */
	jl->j_len = 0;
	atomic_set(&(jl->j_nonzerolen), 0);
	jl->j_start = 0;
	jl->j_realblock = NULL;
	jl->j_commit_bh = NULL;
	jl->j_trans_id = 0;
	jl->j_state = 0;
	put_journal_list(s, jl);
	if (flushall)
		up(&journal->j_flush_sem);
	put_fs_excl();
	return err;
}

static int test_transaction(struct super_block *s,
                            struct reiserfs_journal_list *jl)
{
	struct reiserfs_journal_cnode *cn;

	if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0)
		return 1;

	cn = jl->j_realblock;
	while (cn) {
		/* if the blocknr == 0, this has been cleared from the hash,
		 ** skip it
		 */
		if (cn->blocknr == 0) {
			goto next;
		}
		if (cn->bh && !newer_jl_done(cn))
			return 0;
	      next:
		cn = cn->next;
		cond_resched();
	}
	return 0;
}

static int write_one_transaction(struct super_block *s,
				 struct reiserfs_journal_list *jl,
				 struct buffer_chunk *chunk)
{
	struct reiserfs_journal_cnode *cn;
	int ret = 0;

	jl->j_state |= LIST_TOUCHED;
	del_from_work_list(s, jl);
	if (jl->j_len == 0 || atomic_read(&jl->j_nonzerolen) == 0) {
		return 0;
	}

	cn = jl->j_realblock;
	while (cn) {
		/* if the blocknr == 0, this has been cleared from the hash,
		 ** skip it
		 */
		if (cn->blocknr == 0) {
			goto next;
		}
		if (cn->bh && can_dirty(cn) && buffer_dirty(cn->bh)) {
			struct buffer_head *tmp_bh;
			/* we can race against journal_mark_freed when we try
			 * to lock_buffer(cn->bh), so we have to inc the buffer
			 * count, and recheck things after locking
			 */
			tmp_bh = cn->bh;
			get_bh(tmp_bh);
			lock_buffer(tmp_bh);
			if (cn->bh && can_dirty(cn) && buffer_dirty(tmp_bh)) {
				if (!buffer_journal_dirty(tmp_bh) ||
				    buffer_journal_prepared(tmp_bh))
					BUG();
				add_to_chunk(chunk, tmp_bh, NULL, write_chunk);
				ret++;
			} else {
				/* note, cn->bh might be null now */
				unlock_buffer(tmp_bh);
			}
			put_bh(tmp_bh);
		}
	      next:
		cn = cn->next;
		cond_resched();
	}
	return ret;
}

/* used by flush_commit_list */
static int dirty_one_transaction(struct super_block *s,
				 struct reiserfs_journal_list *jl)
{
	struct reiserfs_journal_cnode *cn;
	struct reiserfs_journal_list *pjl;
	int ret = 0;

	jl->j_state |= LIST_DIRTY;
	cn = jl->j_realblock;
	while (cn) {
		/* look for a more recent transaction that logged this
		 ** buffer.  Only the most recent transaction with a buffer in
		 ** it is allowed to send that buffer to disk
		 */
		pjl = find_newer_jl_for_cn(cn);
		if (!pjl && cn->blocknr && cn->bh
		    && buffer_journal_dirty(cn->bh)) {
			BUG_ON(!can_dirty(cn));
			/* if the buffer is prepared, it will either be logged
			 * or restored.  If restored, we need to make sure
			 * it actually gets marked dirty
			 */
			clear_buffer_journal_new(cn->bh);
			if (buffer_journal_prepared(cn->bh)) {
				set_buffer_journal_restore_dirty(cn->bh);
			} else {
				set_buffer_journal_test(cn->bh);
				mark_buffer_dirty(cn->bh);
			}
		}
		cn = cn->next;
	}
	return ret;
}

static int kupdate_transactions(struct super_block *s,
				struct reiserfs_journal_list *jl,
				struct reiserfs_journal_list **next_jl,
				unsigned long *next_trans_id,
				int num_blocks, int num_trans)
{
	int ret = 0;
	int written = 0;
	int transactions_flushed = 0;
	unsigned long orig_trans_id = jl->j_trans_id;
	struct buffer_chunk chunk;
	struct list_head *entry;
	struct reiserfs_journal *journal = SB_JOURNAL(s);
	chunk.nr = 0;

	down(&journal->j_flush_sem);
	if (!journal_list_still_alive(s, orig_trans_id)) {
		goto done;
	}

	/* we've got j_flush_sem held, nobody is going to delete any
	 * of these lists out from underneath us
	 */
	while ((num_trans && transactions_flushed < num_trans) ||
	       (!num_trans && written < num_blocks)) {

		if (jl->j_len == 0 || (jl->j_state & LIST_TOUCHED) ||
		    atomic_read(&jl->j_commit_left)
		    || !(jl->j_state & LIST_DIRTY)) {
			del_from_work_list(s, jl);
			break;
		}
		ret = write_one_transaction(s, jl, &chunk);

		if (ret < 0)
			goto done;
		transactions_flushed++;
		written += ret;
		entry = jl->j_list.next;

		/* did we wrap? */
		if (entry == &journal->j_journal_list) {
			break;
		}
		jl = JOURNAL_LIST_ENTRY(entry);

		/* don't bother with older transactions */
		if (jl->j_trans_id <= orig_trans_id)
			break;
	}
	if (chunk.nr) {
		write_chunk(&chunk);
	}

      done:
	up(&journal->j_flush_sem);
	return ret;
}

/* for o_sync and fsync heavy applications, they tend to use
** all the journa list slots with tiny transactions.  These
** trigger lots and lots of calls to update the header block, which
** adds seeks and slows things down.
**
** This function tries to clear out a large chunk of the journal lists
** at once, which makes everything faster since only the newest journal
** list updates the header block
*/
static int flush_used_journal_lists(struct super_block *s,
				    struct reiserfs_journal_list *jl)
{
	unsigned long len = 0;
	unsigned long cur_len;
	int ret;
	int i;
	int limit = 256;
	struct reiserfs_journal_list *tjl;
	struct reiserfs_journal_list *flush_jl;
	unsigned long trans_id;
	struct reiserfs_journal *journal = SB_JOURNAL(s);

	flush_jl = tjl = jl;

	/* in data logging mode, try harder to flush a lot of blocks */
	if (reiserfs_data_log(s))
		limit = 1024;
	/* flush for 256 transactions or limit blocks, whichever comes first */
	for (i = 0; i < 256 && len < limit; i++) {
		if (atomic_read(&tjl->j_commit_left) ||
		    tjl->j_trans_id < jl->j_trans_id) {
			break;
		}
		cur_len = atomic_read(&tjl->j_nonzerolen);
		if (cur_len > 0) {
			tjl->j_state &= ~LIST_TOUCHED;
		}
		len += cur_len;